'Models improve where you can measure' — and you can't measure the rest, AI X users argue
teodorio · x · 2026-11-06
An X debate on agent capabilities surfaced a structural point: models systematically improve at measurable tasks and stall at unmeasurable ones. A reply sharpens it further — we don't even know how to measure those unmeasurable abilities, and trying to might actively make things worse. The thread grew out of a jab that people not running agents 10-12 hours a day shouldn't opine on them.
Related event: AI debate: models soar on measurable tasks but stall on unmeasurable ones(2 posts)→
More from AGI Musings
- "Jobs data keeps weakening as AI gets smarter — you can see where this goes" — rand_longevity · 2026-11-06
- The Copenhagen Interpretation of Ethics: the more you try to do good, the more you get blamed — austinc3301 · 2026-11-06
- Quantifying moral tradeoffs is doing good better, even when the math is wrong — austinc3301 · 2026-11-06
- Michael Jordan at ICOMP'26: LLMs are collective intelligence, not a singular super-intelligence — peter_richtarik · 2026-11-06
- NUS professor says Singapore's 30% nationality quota hurts PhD recruiting — xuanalogue · 2026-11-06
- AI's Biggest Cultural Risk: Turning Employees Into the Enemy — DavidLinthicum · 2026-11-06